Alpha Β· in development πŸ‡ΊπŸ‡Ώ Open dataset Dialect-aware

Speech AI for the Uzbek language
that actually understands dialects

UzVoice β€” an open infrastructure for Uzbek speech-to-text and text-to-speech, trained on community-contributed voice data covering all 13 regional dialects of Uzbekistan and Karakalpakstan.

β€”
Contributors
β€”
Submissions
β€”
Hours collected
13
Regional dialects

Existing speech APIs ignore how Uzbeks actually speak

Whisper and Google STT recognize "literary Tashkent Uzbek" with mediocre accuracy and completely fail on Khorezm, Karakalpak, or Surkhandarya dialects. That's not a corner case β€” it's how 60% of the country speaks.

🌍

13 regional dialects

From standard Tashkent to rare Karakalpak β€” each dialect gets dedicated training data and benchmarks.

🎀

Community-driven dataset

Voice contributors are rewarded with points, regional contests, and referral bonuses. Rare dialects earn x3-x5 multipliers.

⚑

Real-time API + on-prem

Sub-500ms streaming STT for call centers, government services, and education. Fully self-hostable for sensitive workloads.

πŸ”’

Open & auditable

Trained on consented data, model cards published, evaluation benchmarks open. No black-box surprises in production.

πŸ“ˆ

Active learning loop

Every API request improves the model: low-confidence audio is routed to human annotators, high-confidence becomes pseudo-labels.

🧠

Trained for Uzbek phonetics

Uzbek-specific G2P rules, fricatives like "q" and "Κ»" properly modeled, code-switching to Russian and Tajik handled.

Dialects we collect β€” and reward

Rare regional dialects get higher reward multipliers in our Telegram collection bot. The rarer the speech, the more valuable the contribution.

Toshkent x1 Andijon x1 Namangan x1 Farg'ona x1 Samarqand x1 Navoiy x2 Sirdaryo x2 Jizzax x2 Buxoro x3 Qashqadaryo x3 Xorazm x4 Surxondaryo x4 Qoraqalpog'iston x5

Roadmap

Public progress: what's shipped, what's in flight, what's next.

βœ“

Telegram collection bot

Live and accepting voice contributions with consent + gamification.

βœ“

Annotation infrastructure

Label Studio + MinIO storage + dialect-aware annotation schema.

βœ“

Auto-transcription pipeline

Whisper large-v3-turbo bootstraps annotators with pre-filled drafts.

●

First 100 hours of dialect data

Targeting May 2026 β€” community contributions + paid annotators.

β—‹

STT alpha API

Public access for select partners (call centers, education, gov).

β—‹

TTS with regional voices

Synthesize speech in any of the 13 dialects, voice cloning for licensed talent.

β—‹

Open dataset release

CC-BY-NC dataset for academic research, partnerships with national universities.

Contribute or partner

πŸŽ™

Donate your voice

Open the Telegram bot, agree to consent, choose your region, and start reading prompts. Earn points, climb regional leaderboards, invite friends.

Open @uzvoicerobot β†’
🀝

Partner / pilot

Looking for early API access, custom dialect models, or full on-premise deployment for your call center, school, or government service?

Get in touch β†’